Robustness Quantification: using imprecise probabilities to assess the reliability of probabilistic classifiers

Reliable prediction is essential in many modern applications of machine learning, yet most classifiers offer little insight into how trustworthy any given prediction truly is. In this seminar, I introduce Robustness Quantification (RQ), a new approach for assessing the reliability of individual predictions by measuring how robust they are to perturbations of the underlying probabilistic model. Drawing on algorithmic techniques from imprecise probabilities, and credal classification in particular, RQ provides computationally efficient methods for quantifying this robustness. Uncertainty Quantification (UQ), on the other hand, focuses on sources of uncertainty rather than on the stability of predictions. I show that RQ remains effective even for small training sets and in the presence of distribution shift, and that RQ and UQ capture complementary aspects of unreliability. Together, they offer a more complete view of when a classifier’s predictions can be trusted and what the underlying sources of unreliability are.